Google’s AI Agent Whitepaper Signals the Dawn of the Agent Era in 2025

Introduction

The whitepaper defines AI agents as autonomous applications that observe the world, use external tools, and plan actions to achieve goals, extending the capabilities of generative language models. It outlines the cognitive architecture consisting of three core components—model, tool, and orchestration layer—and uses a chef analogy to illustrate the agent’s reasoning loop.

Agent Architecture

Agents rely on a language model (LM) as the decision maker, which can be single or multiple LMs, fine‑tuned for specific tasks. Tools bridge the gap between the model and the external world, enabling actions such as database updates, weather queries, or API calls. The orchestration layer manages a cyclic process of information intake, internal reasoning, and action selection until the goal is reached.

Model

The model may be general‑purpose or multimodal and can be fine‑tuned with data that matches the tools used in the cognitive architecture. Prompt examples that demonstrate tool usage can further adapt the model to agent tasks.

Tool Types

Extensions : Standardised API wrappers that let agents invoke external services directly. They are taught through examples and parameters, allowing dynamic selection of the most suitable extension at runtime.

Functions : Model‑generated function calls with arguments that are executed on the client side. This separation gives developers finer control over API execution, security, and data transformation.

Data Stores : Vector‑database backed stores that provide up‑to‑date structured or unstructured information, enabling retrieval‑augmented generation (RAG) and reducing model hallucination.

Reasoning Frameworks

The whitepaper highlights three prompting frameworks that structure the agent’s reasoning:

ReAct – combines reasoning and acting in a single loop.

Chain‑of‑Thought (CoT) – breaks reasoning into intermediate steps.

Tree‑of‑Thoughts (ToT) – explores multiple reasoning paths for strategic tasks.

These frameworks guide the agent to select the next best action, such as choosing a tool from a set like Flights, Search, or Code.

Targeted Learning to Boost Performance

The paper describes three methods for improving an agent’s tool‑use ability:

Contextual Learning : Provide prompts, tools, and a few examples at inference time (e.g., ReAct).

Retrieval‑Based Contextual Learning : Dynamically fetch relevant examples and tools from a data store (e.g., Vertex AI’s example store).

Fine‑Tuning : Pre‑train on a larger, task‑specific dataset to internalise tool usage.

Each method balances speed, cost, and latency, and can be combined for robust solutions.

Practical Example with LangChain

A concrete prototype is built using LangChain and LangGraph with the gemini‑1.5‑flash‑001 model. The agent employs two tools: SerpAPI for web search and Google Places API for location data. Code snippet 8 (illustrated in the accompanying image) shows the workflow, and code snippet 9 displays the resulting multi‑step answer.

Production Deployment on Vertex AI

Vertex AI provides a fully managed environment that bundles the model, tools, orchestration layer, sub‑agents, and example stores. Developers can define agents via a natural‑language UI, test, evaluate, and iterate without handling infrastructure. Figure 15 (image) depicts an end‑to‑end Vertex AI agent architecture, including Vertex extensions, function calls, and example storage.

Summary of Key Takeaways

Agents extend language models with tools to access real‑time information and execute real‑world actions autonomously.

The orchestration layer, powered by reasoning frameworks like ReAct, CoT, and ToT, drives planning and decision‑making.

Extensions, functions, and data stores each offer distinct trade‑offs for tool integration, control, and data freshness.

Targeted learning (contextual prompts, retrieval‑augmented context, fine‑tuning) enhances tool‑use proficiency.

Open‑source libraries (LangChain, LangGraph) and managed platforms (Vertex AI) enable rapid prototyping and production‑grade deployment.

Overall, the whitepaper outlines a modular, iterative approach to building generative AI agents, emphasizing the importance of tool diversity, reasoning frameworks, and continuous experimentation.